Flash-Aware Storage of the Column Oriented Databases

Abstract

Solid state disks become the very popular storage devices. Nonetheless, their architecture based on flash memory has some limitations. They suffer from poor random write performance, as the flash memory blocks must be erased before write. Nowadays, among many database models, the column-oriented databases have attracted the attention. In this model, the data from the particular columns of the database table are stored separately in the memory blocks. As a consequence, only such columns are derived from the memory which are necessary for query execution. In this way, I/O number is reduced, what drastically increases the database performance. This paper proposes a new storage method for the column oriented databases on solid state disks. In this method, the data from each column is stored as a separated structure, called Column-Flash tree (CF-tree). The table is always sorted by a particular column (or a set of columns). Due to utilizing fractional cascading and tree-like structure, the efficiency of update and search is obtained. The storage may be adjusted to the altering query pattern and disk characteristics by changing the height of the CF-tree and the capacity of its levels. In contrast to the other models, the approach may be applied to both read and write optimized databases.

HTAP With Reactive Streaming ETL

“…Conversely, data modification operations (namely INSERT, UPDATE and DELETE operations) need to span multiple pages and hence tend to be more expensive than the equivalent operation in a row-store DBMS. Although, in principle, it is possible to vertically partition a table in a row-store RDBMS to achieve a columnar data organisation, a database engine aware of the columnar data storage can be finely tuned to achieve other benefits, such as improved compression capabilities and late materialization (Abadi et al, 2008;Macyna & Kukowski, 2020), that are not feasible to achieve in a database engine which is tuned for row storage.…”